Divy Yadav

Divy YadavSep 14, 2026

Every AI agent sandbox has to solve the same problem on day one: Get a request from the open internet into an isolated container, without either side trusting the other yet. I spent a few days reading how these 4 platforms actually document that moment. Tensorlake. E2B. Daytona. Fly.io. Same problem. Four genuinely different answers. Some route every request through a full parsing layer (L7). Some just move bytes once a connection is set up (L4). Most platforms mix both, at different hops, for different reasons. Here's what actually stood out to me, though, and it wasn't the architecture. The documentation for three of these platforms gives you a reasonably clear picture of how routing works. Only one went further and published before-and-after measurements from an actual architecture change, down to the specific test conditions. Tensorlake rebuilt a hop in their sandbox proxy from L7 to L4, added kernel TLS, and showed the throughput and CPU cost at every step instead of just claiming it got faster. That's a different kind of claim. You don't have to trust it. You can go check it yourself. I wrote up the full comparison covering the mechanisms, the tradeoffs each platform makes, and four questions I'd now ask about any ingress hop, sandbox or otherwise. Published it n Towards AI, Inc. Link in the comments.

Divy YadavSep 18, 2026

Most RAG tutorials teach you how to make it work. Very few teach you how to make it survive production. You chunk text, run OpenAI embeddings, and dump it into a vector database. It works locally. Then you deploy it and it falls apart. Here is why naive RAG systems break: - Chunks split right down the middle of a thought - Semantic similarity retrieves related text instead of the evidence you actually need - Vector search struggles with multi-hop logic - Zero access control means your bot can expose data users should never see I spent 100+ hours researching, building, and organizing material to understand what actually holds up beyond the notebook. I documented it all in a 137-page curriculum: RAG Engineering: From Basics to Advanced I made this for two kinds of people: Those who want to learn RAG without the noise. And those who think they already know RAG but want to understand the engineering behind it. We cover the real stack: Python • Hybrid Search • Cross-Encoder Reranking • HyDE • Graph RAG • LangGraph • Ragas More importantly, every technique is connected to a specific problem, with clear explanations of when NOT to use it. No hype. Just concrete engineering decisions and code. Get the book: https://lnkd.in/gJC5KcZe What is the hardest bottleneck you've hit when moving RAG out of a notebook?

Divy YadavSep 21, 2026

1,475 subscribers on my newsletter, AI Engineering Simplified And a 50.63% open rate. That number honestly still surprises me every time I check. Building this took longer than I expected. There were stretches where the subscriber count barely moved for weeks. No algorithm boost. No viral moment. Just consistent issues going out, one at a time. The approach was pretty simple: write about things I was actually figuring out. RAG pipelines. Agentic systems. LLM evaluations. Production edge cases. The stuff that doesn't always make the hype threads, but matters when you're actually building. Seems like the right people found it. Current numbers: 1,475 active subscribers 50.63% open rate 4.5% click-to-open rate Still early. But seeing a 50%+ all-time open rate on something I built from scratch in my spare time feels pretty damn good. And I'm just getting started. If you work in AI engineering and want a no-fluff read, the newsletter is here: https://lnkd.in/gvQd5UZV